UltraVR

UltraVR is a diagnostic ultra-resolution image VQA benchmark for evidence-grounded reasoning across remote sensing, CCTV surveillance, and industrial anomaly detection domains.

This repository is a data-only release. It provides benchmark QA annotations, selected redistributable AD images, source mapping files for non-redistributed image domains, and license notices.

Keywords: ultra-resolution image understanding; ultra-high-resolution visual reasoning; evidence-grounded reasoning; visual question answering; vision-language models; multimodal reasoning; remote sensing; CCTV surveillance; industrial anomaly detection.

QA-Only Annotation Release

This trial version keeps the final question, options, answer, question type, one image_path, image dimensions, and license notes. No chain-of-thought is included. In JSONL records, AD image_path values point to local files in this repository, while RS and CCTV image_path values point to the original source-dataset image paths.

Domain Summary

Domain	Source Dataset	Image Files in This Repo	Mapping File	Notes
RS	DOTA-v1.5	No	data/images/rs/mapping.csv	Raw DOTA images are not redistributed.
CCTV	PANDA	No	data/images/cctv/mapping.csv	PANDA is treated as high-resolution still images/screenshots.
AD	MVTec LOCO AD	Yes	Not required	User-provided constructed AD images are included.

Source Datasets

• RS: DOTA-v1.5, https://captain-whu.github.io/DOTA/dataset.html
• CCTV: PANDA, https://gigavision.cn/data/news?nav=DataSet%20Panda&type=nav&t=1781477597958
• AD: MVTec LOCO AD, https://www.mvtec.com/research-teaching/datasets/mvtec-loco-ad

Repository Structure

UltraVR/
├── README.md
├── LICENSE
├── NOTICE
├── RELEASE_POLICY.md
├── DATA_SCHEMA.md
├── data/
│   ├── annotations/
│   │   ├── ultravr_rs.jsonl
│   │   ├── ultravr_cctv.jsonl
│   │   ├── ultravr_ad.jsonl
│   │   └── ultravr_all.jsonl
│   └── images/
│       ├── rs/
│       │   ├── README.md
│       │   └── mapping.csv
│       ├── cctv/
│       │   ├── README.md
│       │   └── mapping.csv
│       └── ad/
│           ├── README.md
│           └── <AD image files>
└── examples/
    ├── sample_rs.jsonl
    ├── sample_cctv.jsonl
    └── sample_ad.jsonl

Annotation Files

All annotations are JSONL files. Each line is one QA sample. data/annotations/ultravr_all.jsonl concatenates RS, CCTV, and AD annotations in that order.

• data/annotations/ultravr_rs.jsonl: 120 QA samples
• data/annotations/ultravr_cctv.jsonl: 120 QA samples
• data/annotations/ultravr_ad.jsonl: 140 QA samples
• data/annotations/ultravr_all.jsonl: 380 QA samples

License

UltraVR follows a mixed research-only licensing policy. Users must follow the original license and usage terms of each source dataset. See LICENSE, NOTICE, and RELEASE_POLICY.md for details.

Citation

If you use UltraVR in your research, please cite our paper:

UltraVR: A Diagnostic Ultra-Resolution Image-VQA Benchmark for Evidence-Grounded Reasoning
Gexin Huang, Yanting Yang, Myeongkyun Kang, Beidi Zhao, Jun Zhou, Chen Zhou, Gang Wang, Zu-hua Gao, and Xiaoxiao Li.
arXiv:2606.05576, 2026.

@misc{huang2026ultravr,
  title         = {UltraVR: A Diagnostic Ultra-Resolution Image-VQA Benchmark for Evidence-Grounded Reasoning},
  author        = {Huang, Gexin and Yang, Yanting and Kang, Myeongkyun and Zhao, Beidi and Zhou, Jun and Zhou, Chen and Wang, Gang and Gao, Zu-hua and Li, Xiaoxiao},
  year          = {2026},
  eprint        = {2606.05576},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2606.05576},
  url           = {https://arxiv.org/abs/2606.05576}
}

Please also cite the original source datasets used by the corresponding UltraVR domains, including DOTA-v1.5, PANDA, and MVTec LOCO AD, when applicable.